KR101871940B1 - Method and system for establishing predictive model of plant abnormality - Google Patents

Abstract

The present invention discloses a method for real-time diagnosis of equipment abnormality having a non-linear relationship with equipment data. To do this, we construct a plant anomaly prediction model by constructing a parameter to construct a plant anomaly prediction model using genetic algorithms. It is possible to diagnose the plant abnormality in real time from the plant data with the finally determined plant abnormality prediction model by repeatedly generating and verifying the model to construct the plant abnormality prediction model until the predetermined standard fitness is satisfied.

Description

[0001] The present invention relates to a method and system for establishing a plant abnormality prediction model,

More particularly, the present invention relates to a method and a system for deriving a relationship model between equipment data and equipment abnormality and diagnosing equipment abnormality from equipment data in real time by applying the model .

The term "equipment data" refers to all data that are calculated while operating and managing the equipment, and "equipment malfunction" means a malfunction or malfunction caused by the equipment itself, excluding external factors that occur when operating the equipment.

In the past, we have defined facility data and equipment abnormalities with clear linearity, and have used various statistical methods to find and manage the relationship. However, when the relationship between the equipment data and the equipment abnormality does not necessarily have a linearity, and there is nonlinearity between the equipment data and the equipment abnormality, it is difficult to find a relationship with the statistical method (multivariate, SPC, PCA) It is difficult to cope with the change, and general facility data does not show uniformity and uniform distribution. Therefore, various methods of non-parametric methodologies have to be found, but it is time consuming and difficult to increase reliability.

The present invention has been made in view of the above problems, and it is an object of the present invention to provide a relational model capable of predicting a plant abnormality from equipment data having non-linearity as well as equipment data having facility abnormality and linearity, .

According to a first aspect of the present invention, there is provided a data management method comprising: a data management step of receiving and synchronizing equipment data; Preparing a model to define the first chromosome; A data dividing step of dividing the facility data into learning data and test data; A model building step of constructing a prediction model reflecting all the learning data from the first chromosome; A model fitness calculating step of calculating the model fitness by inputting the test data into the prediction model; A model determining step of determining the prediction model as a facility anomaly prediction model when the model fitness is higher than or equal to a preset reference fitness; And if the model fit is lower than the preset reference fitness, the current chromosome is newly defined as a new chromosome, and a new prediction model is constructed with the next generation chromosome that crosses and mutates the old chromosome, Repeating the model building repeating step from the step; .

According to another aspect of the present invention, there is provided a data processing apparatus including: a data management unit for synchronizing input facility data; A first chromosomal defining portion defining a first chromosome; A data dividing unit dividing the facility data into learning data and test data; A prediction model construction unit for constructing a prediction model reflecting all the learning data from the first chromosome; A model fitness generating unit for inputting the test data into the prediction model to calculate a model fitness; A fitness comparison unit for comparing the model fitness with a preset reference fitness; And determining the prediction model as an equipment abnormality prediction model when the model fits are higher than or equal to the reference fits. In another case, the present chromosome is newly defined as a global chromosome and the old chromosome is crossed and mutated Generation chromosome definition unit that constructs a new prediction model with a later generation chromosome and repeats the operation until the equipment abnormality prediction model is determined from the operation of the model fitness generation unit.

By using the model derived by the present invention, it is possible to predict a facility abnormality from a plant data with a high probability, even if it has a linear relationship between the plant data and the plant abnormality and has a nonlinear relationship. You can give.

Specifically, by reading the tendency of the facility data and judging the facility abnormality at a short time, it is possible to prevent defects and malfunctions of larger facilities and to prevent accidents in advance.

Fig. 1 shows a conventional method of finding the relationship between equipment data with a clear linearity and a facility abnormality using a statistical method.
FIG. 2 is a block diagram showing a system for predicting a plant anomaly prediction model according to the present invention.
3 is a flowchart showing a construction method of the equipment abnormality prediction model according to the present invention.
FIG. 4 is a flowchart illustrating details of a process occurring in the data management step.
FIG. 5 shows a process of constructing a support vector machine model as an equipment anomaly prediction model, which is one of embodiments of the present invention. In particular, FIG. 5 is a flowchart showing operations in a model construction preparation unit and a model construction unit excluding a data management unit.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings. It should be noted that the same components of the drawings are denoted by the same reference numerals and signs as possible even if they are shown on different drawings.

In the following description of the present invention, a detailed description of known functions and configurations incorporated herein will be omitted when it may make the subject matter of the present invention rather unclear. On the other hand, when an element is referred to as "comprising ", it means that it can include other elements as well, without excluding other elements unless specifically stated otherwise.

Fig. 1 shows a conventional method of finding the relationship between equipment data with a clear linearity and a facility abnormality using a statistical method.

And out of facility management data FAULT I to IV out of the normal management line.

Statistical Classification Model is used to find the correlation of two data. Classification models can be constructed by machine learning algorithms. By inputting given data into a computer and constructing a discrimination criterion by performing learning based on a certain algorithm, new types of data are given, And the process of predicting whether it is discriminated. Machine learning algorithms include K-Nearest Neighbors algorithm, Perceptron, Radial Basis Function Network, Genetic Algorithm (GA), Support Vector Machine (SVM) Support Vector Machine) and the like. In the present invention, it is proposed to construct an equipment abnormality prediction model through at least one of the above algorithms. The equipment abnormality prediction model of the present invention means a model capable of diagnosing equipment abnormality by receiving equipment data. In particular, one of the embodiments of the present invention uses a genetic algorithm and a support vector machine at the same time.

Genetic Algorithm is a kind of machine learning algorithm that is based on Darwin's theory of evolution. It is an algorithm that is performed in a way that evolves itself according to the surrounding environment.

Genetic algorithms can be developed by making multiple chromosomes according to one of several selection methods, distinguishing the superiority of each chromosome, removing the remaining chromosomes and removing the remaining chromosomes, It will continue to evolve until it comes out. A detailed description thereof will be described later with reference to FIG.

The support vector machine is a method of predicting new data by learning the discrimination method of dividing two kinds of data properly by computer. The support vector machine model has a limitation that it can not achieve satisfactory performance in the real problems of nonlinear classification. However, it can be effectively predicted by applying the mapping method using a function called kernel, Mapping refers to moving the nonlinear problem, which is difficult to divide in the input space where we actually place the data, to the high dimensional space called the feature space and performing the linear discrimination of the support vector machine in this new space, Indicates that the same effect as solving the complex nonlinear discrimination problem in the input space is obtained.

To use the support vector machine model, we need two input parameters, which are used to determine the input vector, and kernel function parameters, to use the kernel function. A detailed description thereof will be given later with reference to FIG. 5, and since the support vector machine described above is a known technology, a general description not directly related to the present invention is omitted for the sake of simplicity of the invention.

FIG. 2 is a block diagram showing a system for predicting a plant anomaly prediction model according to the present invention.

2, the equipment abnormality prediction model construction system according to the present invention includes a data management unit 200, a model construction preparation unit 210, and a model construction unit 220.

The data management unit 200 manages facility data and includes a data input unit 201 and a data completion unit 202. The data input unit 201 reads and stores the production data file and the trace data file in the form of a text file and synchronizes them. The data completion unit 202 sets a reaction variable indicating an equipment abnormality with an accumulated loss rate, and completes the equipment data by including the response variable itself in the production data. That is, the completed facility data includes not only the data at the time when the equipment normally operates in the past, but also the data at the time when the equipment abnormality occurs. The completed synchronized facility data is transmitted to the model building preparation unit 210.

The model construction preparation unit 210 finds a parameter for constructing an optimal equipment abnormality prediction model through a genetic algorithm and generates a first chromosome definition unit 211, a fitness comparison unit 212, and a next generation chromosome definition unit 213 ).

The first chromosome defining unit 211 defines a first chromosome, which is a variable for constructing a 'prediction model' experimentally created to determine an 'equipment abnormality prediction model'. The number of construction parameters of the prediction model included in the first chromosome may be plural as well as the number of stages. For example, according to the embodiment of the present invention in which the prediction model is a support vector machine model, And kernel function parameters. The fitness comparison unit 212 and the later generation chromosome definition unit 213 will be described later.

The first chromosome defined in the first chromosome defining unit 211 is transmitted to the model building unit 220 together with the completed equipment data in the data completion unit 202.

The model construction unit 220 includes a data division unit 221, a prediction model construction unit 222, and a model fitness generation unit 223.

The data division unit 221 receives the facility data completed by the data completion unit 202 and divides the facility data into learning data and test data.

The predictive model building unit 222 constructs a predictive model using the first chromosome or the next generation chromosome received from the model construction preparation unit 210 and performs a learning operation so that all of the learning data of the data division unit 221 can be reflected . Generation chromosome will be described later in the chromosome definition unit 213 of the later generation.

The model fitness generator 223 applies the test data to the 'prediction model' received from the prediction model building unit 222 to generate a model fitness.

The fitness comparison section 212 of the model preparation preparation section compares the model fitness generated by the model fitness generation section 223 with a preset reference fitness.

The next generation chromosome definition unit 213 of the model construction preparation unit sets the currently defined chromosome as a global chromosome when the model fitness is lower than the standard fitness in the fitness comparison unit 212 and sets the cross chromosome as a crossing and mutation according to the preset ratio To define a new generation of chromosomes. If the model fitness is higher than or equal to the reference fitness, the current forecast model is determined as the equipment anomaly prediction model.

3 is a flowchart showing a construction method of the equipment abnormality prediction model according to the present invention. Each step of Fig. 3 is performed by the equipment abnormality prediction model construction system of Fig.

In step S300, facility data is received and synchronized to utilize the facility data. This will be described later in detail with reference to FIG.

In step S400, a first chromosome composed of variables for constructing a prediction model is defined. The process other than defining the initial chromosome will be described later in step S700.

In step S500, a prediction model is constructed from the first chromosome. Specifically, the step of dividing the experimental data into the learning data and the test data should be performed prior to the construction of the prediction model, and the prediction model constructed from the initial chromosome is subjected to the learning process by the learning data among the divided experimental data. A detailed description of the learning process will be described later with reference to FIG.

In step S600, test data is input to a predictive model that has undergone a learning process to derive an equipment abnormality based on test data, and a model fitness is calculated therefrom.

In step S700, a process of comparing the model fitness of step S600 with a preset reference fitness is shown. Criteria fitness is a measure of the need to stop the iterative process of constructing a predictive model, which consists of false positives and coverage.

The term "false positive" refers to a rate at which a positive result is obtained, which is not actually positive. In the first embodiment of the present invention, "90%" is applied, and this value can be changed according to the embodiment.

The coverage indicates the degree to which the prediction model is determined to be abnormal in the equipment. In the first embodiment of the present invention, '30%' is applied, and this value can be changed according to the embodiment.

For example, according to the first embodiment of the present invention, when an equipment abnormality number 100 occurs, the equipment abnormality prediction model should report to the facility manager that at least 30 times of equipment abnormality has occurred due to coverage, When a facility manager reports to the facility manager that there is an equipment malfunction, the equipment malfunction should have occurred at least 27 times.

If the model fitness is greater than or equal to the reference fitness, the process proceeds to the model determination phase. Otherwise, the process proceeds to the model construction preparation phase.

When proceeding to the preparation stage of the model, the current chromosome is defined as the former chromosome, and then crossing and mutation are applied to generate the next generation chromosome. Subsequent chromosomes become the current chromosomes on behalf of the former chromosomes, and in detail, are the variables that can be used to construct new prediction models.

In step S800, a prediction model showing a model fitness that is higher than or equal to the reference fitness is determined as the equipment abnormality prediction model. The determined equipment abnormality prediction model is a model capable of diagnosing equipment abnormality with respect to all equipment data inputted in step S300 by more than the standard fitness.

FIG. 4 is a flowchart showing details of a process occurring in the data management step.

In step S310, the production data file and the trace data file in the form of a text file are read, stored, and synchronized. The production data is the data produced by the facility, and the data for correlating with the production data is trace data.

In step S320, a response variable is set with an accumulated loss rate obtained by digitizing the product to be discarded. The cumulative rate of abandonment is a cumulative value obtained by repeating the instantaneous rate of abandonment, which is the rate of discarding the product when the facility abnormality occurs. The higher the response variable is calculated, the higher the response variable is the frequent abnormality of the facility. The response variable can be measured by the defect rate, the mounting error, the cumulative rejection rate, etc. In the present invention, the reaction variable is calculated by the cumulative rejection rate which is easy to know the tendency.

In step S330, the completed synchronized facility data is transmitted to the model building preparation unit. The completed synchronized facility data includes response variables determined by production data, trace data and cumulative rejection rates.

The completed synchronized equipment data includes not only the data at the time when the equipment normally operates in the past but also the data at the time when the equipment abnormality has occurred. Therefore, in the equipment abnormality prediction model finally built based on this data, When the same facility data as the facility data causing the abnormality is inputted, it is possible to diagnose the facility abnormality irrespective of the relation between the facility data and the facility abnormality.

FIG. 5 is a flowchart showing a process of constructing a support vector machine model as a plant anomaly prediction model, which is one of embodiments of the present invention, and particularly, operations in a model construction preparation unit and a model generation unit excluding a data management unit.

In step S401, a combination of input variables and a kernel function parameter for generating a support vector machine are selected. This is the first step to proceed with the genetic algorithm, and these two elements constitute the chromosome to be described later in step S403.

The combination of input variables is a value indicating the size and direction of the input vector to be applied to the support vector machine. The kernel function is a function that allows the support vector machine model to perform effectively between two nonlinear data, including a linear kernel, a polynomial kernel, a radial basis kernel (RBF) Basis Function kernel). In each kernel, there are parameters that help optimization. There is no automatic way to find out which parameter is best to choose, so it is obtained through iterative process of genetic algorithm.

The selection in the genetic algorithm is a step of determining a plurality of initial solutions to be input in order to obtain a desired final solution, and there are various methods such as roulette wheel selection, tournament selection, rank based selection, and the like. If the initial solutions are appropriately determined, the iterative process specific to the genetic algorithm is minimized, and the final solution can be obtained in a short time. For example, the initial solution may be selected as (15,04), (13,07), (11,10), (09,13). The order is assumed to be a combination of input variables and kernel function parameters.

In step S402, a combination of the input variables and a binary number are converted. (1111, 0 100), (1101, 0111), (1011, 1010), (1001, 1101) are obtained by using the solution selected in step S402.

In step S403, the combination of the input variables and the kernel function parameters are defined as first chromosomes. The chromosome is the input unit to drive the genetic algorithm, and the chromosome initially determined will evolve to produce the final chromosome composed of the final solution. If the solution selected in step S403 is used, the initial chromosomes are 11110100, 11010111, 10111010, and 10011101, respectively.

In step S404, the facility data transmitted from the first chromosome and the data management unit in Fig. 2 are transmitted to the model building unit.

In step S405, the facility data received from the model building preparation unit via the data management unit is divided into the learning data and the test data. The partition ratio is not fixed, but for the validity of the iterative process of the genetic algorithm, the value of either learning data or test data should not be zero, and the learning data and test data should be facility data including equipment . Since the process of dividing the facility data into the learning data and the test data does not affect the other steps, it can be performed not only by the model building unit but also by the model building preparation unit.

In step S406, a support vector machine model that reflects the contents of the learning data to the current chromosome is constructed.

Here, the current chromosome may be the first chromosome defined in step S404, or a later generation chromosome defined by step S409 described later. This will be described later in step S409.

Reflecting all the contents of the learning data means that the support vector machine model constructed with the current chromosome should be a model that can predict the equipment abnormality more completely than the learning data except the test data.

Since the learning data includes not only the data related to general facility operation but also the equipment abnormal data, it is possible to judge whether the support vector machine model constructed with the current chromosome operates only with the learning data. A support vector machine model constructed by only a combination of input variables for classifying is used as a first predictive model and a first predictive model is learned as learning data, Can be called.

In step S407, test data is input to the constructed support vector machine model to calculate the model fitness, and the model fitness is transmitted to the model construction preparation unit together with the support vector machine model.

Since the support vector machine model is optimized only for the learning data, there is no guarantee that the equipment error will be completely predicted for the test data. Therefore, the test data is input to the current support vector machine model to calculate the equipment abnormality data. Then, the model anomaly is further calculated by comparing the equipment abnormality data for the test data predicted by the support vector machine model with the equipment abnormality data based on the past test data contained in the test data.

The model fitness is a measure of how well the established support vector machine model can diagnose faults over the plant and is calculated for comparison with a pre-established reference fitness. The high fitness of the model means that not only the learning data but also the test data can be predicted according to the standard set by the facility manager.

Finally, the model fits calculated for comparison with the reference fitness and the support vector machine model generated by the model building unit are transmitted back to the model construction preparation unit.

In step S408, the predetermined reference fitness is compared with the model fitness calculated in step S407.

In step S409, it is determined whether to continue the prediction model construction based on the comparison of the reference fitness and the model fitness.

If the model fitness does not satisfy at least 90% of the false positive and 30% or more of the coverage, the next generation chromosome is defined and the process proceeds to step S406.

The next generations chromosome refers to the chromosomes obtained by applying the crossover and the mutation to the chromosome used to construct the current support vector machine model as the former chromosome.

In the case of the intersection, a two-point crossing, a three-point crossing, and the like can be used, and the probability of occurrence of the mutation can be suitably set to a low value so as not to interfere with the convergence of the final solution.

For example, if a pair of first chromosomes are 11110100 and 11010111, as exemplified in step S403, and the two-point crossover occurs at positions 2 to 3 and 7 to 8, respectively, the chromosome of the next generation becomes 11010110. Here, when the mutation occurs at the 8th position, 11010111 is identified as the next generation chromosome.

The confirmed next generation chromosome becomes the current chromosome and builds a new support vector machine model in step S406 and repeats the later steps until it calculates the model fit above the reference fit. If the genetic algorithm is sufficiently repeated and the model fidelity of the predicted model that is calculated is significantly lower than the standard fidelity, the process returns to step S405 instead of step S406, and the division ratios or the division configurations of the facility data may be different.

If the model fit satisfies both false positive 90% and coverage 30% or more, the current support vector machine model is determined as the equipment anomaly prediction model. According to the concept described above in step S406, when the model goodness of the second prediction model is higher than the reference goodness, the second prediction model is determined as the equipment anomaly prediction model.

Even if the model constructed by the present invention shows a nonlinear relationship between equipment data (production information of equipment, event information, sensor information, and the like) and equipment abnormality, the equipment abnormality can be estimated with high probability from the equipment data in real time. And it is excellent in real-time diagnosis of mounting abnormality of surface mounting technology (SMT) equipment with a clear tendency of cumulative rejection rate.

The present invention is not limited by the above-described embodiments and the accompanying drawings, and the present invention can be implemented by a machine learning algorithm other than the support vector machine described above. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims. It will be self-evident.

200:
201: Data input unit
202: Data completion unit
210: Preparing the model
211: First chromosome definition part
212: fitness comparator
213: Next generation chromosome definition part
220: Model construction unit
221:
222: prediction model construction unit
223: Model fitness generator

Claims (6)

A data management step of inputting and synchronizing equipment data of a facility for producing a product for each process;
Preparing a model to define the first chromosome;
Dividing the synchronized facility data into learning data and test data;
A model building step of constructing a prediction model reflecting all the learning data from the first chromosome;
A model fitness calculating step of calculating the model fitness by inputting the test data into the prediction model;
A model determining step of determining the prediction model as a facility anomaly prediction model when the model fitness is higher than or equal to a preset reference fitness; And
If the model fit is lower than the predetermined criterion fitness, the current chromosome is newly defined as a global chromosome, a new prediction model is constructed with a next generation chromosome that crosses and mutates the old chromosome, And repeating the model constructing and repeating steps from the step
The data management step comprises:
Wherein the facility includes an accumulated loss rate calculated based on the amount of the discarded product in the synchronized equipment data when discarding the product produced through the process in which the equipment abnormality occurs,
Wherein the equipment abnormality prediction model predicts an abnormality operation time point of the facility based on the accumulated discard ratio included in the input new equipment data,
The reference fidelity,
False positive and coverage. ≪ RTI ID = 0.0 > 11. < / RTI > The method of claim 1,
And a support vector machine (SVM) model is used. 3. The method of claim 2,
The first chromosome, the old chromosome and the next generation chromosome,
A combination of input variables represented by binary numbers and a kernel function parameter. 4. The method according to any one of claims 1 to 3,
Wherein the facility abnormality and the facility data have a mutually non-linear relationship. The method according to claim 1,
And a facility abnormality prediction model for diagnosing a facility abnormality from equipment data flowing into the facility abnormality prediction model in real time. A data management unit for receiving and synchronizing facility data of a facility that produces a product for each process;
A first chromosomal defining portion defining a first chromosome;
A data dividing unit for dividing the synchronized facility data into learning data and test data;
A prediction model construction unit for constructing a prediction model reflecting all the learning data from the first chromosome;
A model fitness generating unit for inputting the test data into the prediction model to calculate a model fitness;
A fitness comparison unit for comparing the model fitness with a preset reference fitness; And
The prediction model is determined as the equipment anomaly prediction model when the model fitness is higher than or equal to the criterion fitness, and in the other cases, the current chromosome is newly defined as a global chromosome, and the old chromosome is crossed and mutated Generation chromosome definition unit that constructs a new prediction model with a later generation chromosome and repeatedly operates from the operation of the model fitness generation unit to the determination of the equipment abnormality prediction model,
The data management unit,
Wherein the facility includes an accumulated loss rate calculated based on the amount of the discarded product in the synchronized equipment data when discarding the product produced through the process in which the equipment abnormality occurs,
Wherein the equipment abnormality prediction model predicts an abnormality operation time point of the facility based on the accumulated discard ratio included in the input new equipment data,
The reference fidelity,
A false positive and a coverage. The system of claim 1,

Images